These replacement rules assume that TST and TEQ set V. But TST and
TEQ do not set V. This is a problem because instructions like LT are
actually checking for N!=V. But with TST and TEQ not setting V, LT
doesn't do anything meaningful. It's possible to construct trivial
miscompilations from this, such as:
package main
var x = [4]int32{-0x7fffffff, 0x7fffffff, 2, 4}
func main() {
if x[0] > x[1] {
panic("fail 1")
}
if x[2]&x[3] < 0 {
panic("fail 2") // Fails here
}
}
That first comparison sets V, via the CMP that subtracts the values
causing the overflow. Then the second comparison operation thinks that
it uses the result of TST, when it actually uses the V from CMP.
Before this fix:
TST R0, R1
BLT loc_6C164
After this fix:
TST R0, R1
BMI loc_6C164
The BMI instruction checks the N flag, which TST sets. This commit
fixes the issue by using [LG][TE]noov instead of vanilla [LG][TE], and
also adds a test case for the direct issue.
Fixes#42876.
Change-Id: I13c62c88d18574247ad002b671b38d2d0b0fc6fa
Reviewed-on: https://go-review.googlesource.com/c/go/+/274026
Run-TryBot: Jason A. Donenfeld <Jason@zx2c4.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Cherry Zhang <cherryyz@google.com>
Trust: Jason A. Donenfeld <Jason@zx2c4.com>
